moreyash007/nerofence

NeuroFence is an air-gapped desktop scanner for detecting LLM weight poisoning and backdoors. An analyst selects a local model; the worker audits and safely loads it, records baseline activations, exercises adversarial prompts, and scores anomalies before returning a report to the desktop console. A separate static viewer reads exported reports locally and presents scores, flagged neurons, integrity findings, layer profiles, and activation heatmaps. The README documents these workflows; sampled source confirms key orchestration and viewer paths, while some subsystem wiring is documented rather than directly sampled.